Reflections after 30 days of using GitHub Copilot Agent.

vibe-coding

Reflections after 30 days of using GitHub Copilot Agent.

My vibe coding project: live App Agentic AI Apps. And code repo -> Agentic AI Apps Repo

This project is a demonstration and experimental modern application of AI agents. It features a variety of agent types, including Instruction Agents, Function Call Agents, Workflow Agents, and MCP Agents.

The project contains over 60 code files and uses a distributed deployment approach. The entire application development process was driven by me providing requirements in natural language, while GitHub Copilot Agent handled the coding.

Technology Stack:

  • Framework: Next.js, Typescript, Python, Tailwind CSS, Copilotkit, Langchain
  • AI Models: GPT-4o, GPT-4o-mini, GPT-o3-mini, Deepseek v3, Deepseek-r1 from GitHub Models and Azure OpenAI
  • Deployment: Azure Web Apps, Azure Function, Azure container registry, GitHub Action, Supabase

Transformations We Must Embrace Today:

  • Embrace abstraction rather than getting lost in details.
  • Focus on solving problems rather than controlling every aspect.
  • Pursue continuous iteration rather than achieve success in one go.
  • Achieve rapid usability rather than striving for flawless polish.

Learn from the journey

  • Precise Descriptions: Clear and accurate prompts are essential, as is dev framework terminology.
  • Adopt Modern Frameworks: Prioritize modern frameworks and adhere to the principles of low coupling and high cohesion. Limit each file to no more than 1,000 lines.
  • Agent or Edits: As the project gets larger, it is actually the selection file that carries the Edits more than the Agent for the entire codebase.
  • Consistency with Rollback Strategies: Inconsistencies may arise in larger projects, making ‘undo’ functionality and strong rollback strategies indispensable.
  • Start from Workable: Version control is important, keep a workable version before iterating.
  • Context Matters for AI Agents: File names, comments, strings, variables, and function names all serve as critical context for AI-driven programming.
  • Agent Wrong or Human Wrong?: The model keeps increasing the check output when the run fails, but it is possible that the problem lies elsewhere.
  • Provide Sufficient and Useful Example: Providing Agent with well-designed examples and meaningful references greatly improves their performance.
  • Controlled Refactoring by Agents: Be cautious when assigning refactoring tasks to agents. Clearly define and limit the scope of changes.
  • Know When to Let Go: Recognize when abandoning an unproductive approach is more efficient than persisting with repeated attempts.
  • Use AI to Test AI: Let the agent do the testing, unit test yes, UI test also.
  • “It Works” is Top Priority: Prioritize continuous deployment and functional utility over well prepare. If it works, it’s good enough.
  • Avoid Path Dependency: Vibe Coder to try to avoid path dependency in thinking and methodology.
  • Free Yourself: Vibe Coder should avoid getting caught up in unnecessary details or overly dwelling on insignificant matters.

===

今天我们需要做出改变：

  • 理解抽象，而不是沉溺细节
  • 解决问题，而不是掌控全程
  • 不断迭代，而不是一蹴而就
  • 快速可用，而不是完美打磨

Vibe coding 心得体会

  • 精准描述需求的提示词仍然很重要，开发和框架术语也很重要。
  • 尽量使用现代的框架，低耦合高内聚原则，每个文件不要超过1000行。
  • 项目规模变大后，实际是选择文件进行Edits要多于整个codebase的Agent。
  • 一致性可能会出现问题，undo很重要，做好roll back策略。
  • 版本控制很重要，保持一个可用版本，再进行迭代。
  • 面向AI Agent 编程，文件名，注释，字符串，变量名，函数名皆是上下文。
  • 运行失败时，模型会不断增加check输出，但有可能问题出在其他方面。
  • 给模型提供参考和示例也同样重要。
  • 要求Agent重构需谨慎，范围需要控制。
  • 及时放弃有时候比不断尝试更重要。
  • 让agent来做测试，unit test是，UI test也是。
  • Vibe coding，持续部署，能用就是好的。
  • Vibe Coder要尽量避免思维和方法论上的路径依赖。
  • Vibe Coder要克制自己陷入细节和计较无所谓的东西。